// the find
marqo-ai/marqo
Ecommerce Search and Discovery - marqo.ai
Marqo is a vector search platform originally aimed at ecommerce product discovery — multi-modal (text + image) search built on top of Vespa and custom embedding pipelines. The open-source version is now officially deprecated; the team has moved everything to a managed SaaS product. This repo is a graveyard.
The inference orchestrator is well-structured — clean separation between model management, embedding pipelines, and the Triton inference backend. The LFU/LRU inference cache implementation is thoughtful, with monitoring hooks baked in. CI coverage is serious: separate workflows for AMD64, ARM64, CPU, CUDA, shards/replicas, and backwards compatibility. The multi-modal architecture (text + image in the same index via OpenCLIP) was genuinely ahead of most open-source search projects when it launched.
The README leads with a deprecation notice — this is not a codebase you should adopt. Active development has stopped and bugs will not be fixed. The operational footprint is heavy: you need Vespa, a Triton inference server, and the orchestrator running together before you can index anything, which makes local development painful. No self-hosted upgrade path is offered; the deprecation pushes you to their paid cloud, not to a community fork. Documentation links point to docs.marqo.ai which will likely drift or disappear as the SaaS product diverges.